The agent answers, the CLI does the CSS

What it does

Answer me with HTML is an agent skill for Claude Code, Codex, Cursor, OpenCode and 70-odd other agents. When you ask a question that deserves more than a paragraph — a protocol walkthrough, a repo’s module map, a Redis-vs-Memcached call — the agent writes a short Markdown draft and hands it to a bundled CLI, which renders it into a single styled HTML page with diagrams, tables and timelines. The agent decides on its own when a page is worth it; a one-line question still gets a one-line answer.

The interesting bit

The trick is token economics. Instead of the model typing every div, CSS rule and SVG coordinate, it writes only content — about 1/7 of the tokens. The README’s own benchmark (3 topics × 3 runs, Claude Sonnet 5.5, medians) shows 923 output tokens vs 6,873 for asking for HTML directly, and 13 s vs 46 s. Cost stays roughly flat because the skill adds turns that re-read context — you save the waiting, not the bill. There’s also a 3Blue1Brown-style video mode: the same draft plus one narration line per beat becomes a self-contained player page where arrows draw themselves and [Server] in the narration pans the camera to that node.

Key highlights

• 7.4× fewer output tokens and 3.6× faster than asking the model for raw HTML, per the bundled benchmark
• Diagrams, folder trees, call graphs, comparison tables and timelines generated from short Markdown drafts
• Explainer videos with spoken narration (ElevenLabs, system voice, or a local OpenAI-compatible TTS server), all in one offline-playable file
• Optional always-on mode attaches a small 2–4 panel page to every conclusion, without popping the browser
• No config files to edit — settings via slash commands or am config; runs fully offline apart from an optional weekly version check

Caveats

• The benchmark is small (3 topics, 3 runs) and self-published; treat the exact multipliers as indicative
• Video export to MP4 needs Chrome, ffmpeg and Node 22+ on your machine
• Always-on mode adds a ~90-token reminder to every turn, which is a real (if small) context tax

Verdict

Worth installing if you regularly ask agents to explain complex systems and are tired of either wall-of-text answers or waiting 46 seconds for hand-written HTML. Skip it if your questions are mostly one-liners — the agent will (correctly) ignore it anyway.